Abstract
Purpose:
To identify retinal microvascular features that distinguish normoglycemia from diabetes mellitus (DM)[Study-1] and prediabetes (PreDM)[Study-2] among Asian Indians using artificial intelligence (AI); evaluate their diagnostic accuracy; and examine independent associations of oculomics scores [Oculomic Diabetes Score (ODiS) and Prediabetes Score (OPreS)] with glycemic status.
Methods:
We analyzed 273 retinal images from 139 participants (19 = normoglycemia, 100 = DM, 20 = PreDM; mean age: 49.4 ± 12.9 yrs; male: 71.2%) with dataset randomly split (50:50) into training and test sets. We extracted 226 quantitative vessel tortuosity features separately for arteries and veins using machine vision-based approaches. Top six discriminating features were separately selected (using Wilcoxon Rank-Sum test and Linear Discriminant Analysis) on training sets and validated on independent blinded test sets. Model performances were evaluated using area under precision-recall curve (AUPRC) for DM and area under the receiver operating characteristic curve (AUC) for PreDM. Independent association of ODiS and OPreS (adjusting for age, sex, height, weight, blood pressure, serum cholesterol, serum creatinine) was assessed by multivariable logistic regression.
Results:
Specific oculomics-based retinal vascular features distinguished DM/PreDM from normoglycemia. Vein-only model achieved AUPRC = 0.96 (95% confidence interval [95% CI]: 0.9–1.00) (sensitivity = 95%, specificity = 72.2%, precision = 95%) on Sv1 for DM and an AUC = 0.80 (95% CI: 0.63–0.94) (sensitivity = 70%, specificity = 80%, precision = 82.35%) on Sv2 for PreDM. Oculomics scores were independently associated with DM(ODiS) [Adjusted Odds ratio (AdjOR): 2.00 (95% CI: 1.56–2.57, P < 0.001)] but not with PreDM(OPreS) [AdjOR: 1.32 (95% CI: 0.65–2.71, P = 0.48)].
Conclusions:
In this proof-of-concept study, AI-informed retinal venous features on routine fundus images, with further prospective and multisite validation, could potentially serve as noninvasive DM detection using AI models among Asian Indians.
This is a visual representation of the abstract.
Get full access to this article
View all access options for this article.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
